HAL Id: hal-03202349
https://hal.archives-ouvertes.fr/hal-03202349
Submitted on 27 Apr 2021
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Riemannian Geometry on Connectivity for Clinical BCI
Camille Noûs, Marie-Constance Corsi, Sylvain Chevallier, Florian Yger
To cite this version:
Camille Noûs, Marie-Constance Corsi, Sylvain Chevallier, Florian Yger. Riemannian Geometry on Connectivity for Clinical BCI. ICASSP 2021 - IEEE International Conference on Acoustics, Speech and Signal Processing, Jun 2021, Toronto / Virtual, Canada. �hal-03202349�
Riemannian geometry on connectivity for clinical BCI – RIGOLETTO
A participation to the IEEE WCCI BCI Clinical Challenge
1. ARAMIS LAB, Inserm U1127, CNRS UMR 7225, UPMC, ICM, Inria, Paris, France
2 LAMSADE, CNRS, Univ. Paris-Dauphine, PSL Research University, Paris, France 3 LISV, Univ. Paris-Saclay, Versailles, France
4 Cogitamus, CNRS, Paris, France
Marie-Constance Corsi 1 Florian Yger 2 Sylvain Chevallier 3
Controlling a brain-computer interface (BCI) requires time to achieve high performance [1]. Despite its clinical applications, one of the main drawbacks is the high inter-subject variability that could be noticed for performance. This is sometimes referred in the literature as the « BCI inefficiency » phenomenon [2] and affects its usability. Among the approaches adopted to tackle these issues are the search for neuromarkers, that potentially capture better the neurophysiological mechanisms underlying the BCI performance and the optimization of the classification pipelines, that could be robust enough to be applied to any subject.
In this work, we proposed an original approach that combines functional connectivity estimators, Riemannian geometry and ensemble learning to ensure a robust classification. This methods was ranked First in the Clinical BCI Challenge-WCCI (within-subject category).
The authors would like to thank the organizers of the WCCI Clinical BCI competition for giving them the opportunity to kickstart this promising collaboration. FY thanks F. Lotte for suggesting several years ago the use of connectivity features in lieu of covariances in Riemannian classifiers.
Acknowledgements
To assess the performance of our approach we conducted a comparison of the estimators both at the individual and at the group level. For that purpose, we directly applied our approach to an EEG dataset obtained from 8 stroke patients, in the frame of the IEEE WCCI
Proof of concept
Our approach got the first position on this task with a substantial margin, the following teams having respectively kappa scores of 0.49 and 0.47 and accuracies of 74.69% and 73.75%. The kappa score obtained on validation is close to 0.68 and to the value obtained on training data with a 5-fold cross-validation. Our hypothesis is that this result is due to the
Functional connectivity
Two main estimates were computed within the alpha-beta frequency band (8-30Hz): the coherence (Coh) and the phase-locking value (PLV).
We also computed two other FC estimates: the amplitude envelope correlation (AEC) and the imaginary coherence (ICoh). For the sake of completeness, we report the results of AEC and ICoh although those features were not used in the final submission. For an illustrative purpose, we represented the values obtained by averaging the FC estimators in each task separately in data obtained from one patient.
Contact information:
@MConstanceCorsi
[1] J. Wolpaw, et al, “Brain–computer interfaces for communication and control,” Clinical Neurophysiology, 2002.
[2] M. Thompson, “Critiquing the Concept of BCI Illiteracy,” Science and Engineering Ethic, 2018.
[3] F. Lotte, et al, “A Review of Classification Algorithms for EEG-based Brain-Computer Interfaces: A 10-year Update,” Journal of Neural Engineering, 2018.
References
Riemannian geometry
The diffusion of Riemannian geometry for motor imagery BCI has strongly increased in the past decade [3]. This method consists in the extraction of Symmetric Positive Definite (SPD) matrices, usually the covariance matrices among sensors, for each trial and in considering this as a curved space.
Here, to estimate the distance between two matrices A & B, we used the LogEuclidean distance via the Froebenius norm:
𝛿!" 𝐴, 𝐵 = log 𝐴 − log(𝐵) 𝔉
Coherence (Coh) Phase-Locking value (PLV)
Df(t), relative phase between s1 and s2
s1(t) s2(t) !"ℎ!"(%) = (!"(%) " (!!(%)(""(%) )*+ = ,#∆%(') S12(f), cross-spectral density S11(f), auto-spectral density
Correlation Coherence Phase Locking Value
Left
Right
Correlation Coherence Phase Locking Value
Left
Right
Correlation Coherence Phase Locking Value
Left
Right
Spatial covariance (Cov), coherence (Coh) and phase-locking value (PLV) are estimated from electroencephalographic (EEG) signals. A first level of classification was performed by FgMDM classifiers that yielded output decision probabilities to train a second level classifier, a ridge regression classifier, that provided the final decision. The associated code is available at https://github.com/sylvchev/wcci-rgcon.
Ensemble classifier
EEG
Cov Coh PLV FgMDM FgMDM FgMDMRidge
classifier
Decision
Clinical BCI Challenge. There were 14 submissions to the competition from 12 different institutions around the world across 9 different countries spread across 3 continents.
Using FC estimators associated with an ensemble classifier gives the possibility to take into account the users’ specificity. After participating to the competition, we identified different approaches to improve our method. Further investigation should be done regarding the follow-up to the MDM and the dimensionality reduction.. We plan to refine the selection of the frequency band of interest. Another promising lead would be to extract for each epoch several PSD matrices, each on a different frequency band, and to consider this set as a trajectory on the manifold.
robustness provided by a classification based on multiple sources.
